But, how do I calculate risk-adjusted metrics, what are they?
#mathishard
W-Score
The W-score quantifies how a trauma center performs relative to expected outcomes.
It expresses the difference between observed and expected survivors, scaled to patient volume.
\[
W = \frac{A - B}{C} \times 100
\]
W-Score
Where:
\(A\) = Total number of patients with all data necessary to calculate \(P(Survival)\)minus the number of those patients who died
\(B\) = Sum of all predicted survival probabilities \(P(Survival)\) for this patient group
\(C\) = Total number of patients with all data necessary to calculate \(P(Survival)\)
W-Score
Interpretation for clinicians:
\(W > 0\) -> More survivors than expected; center performing better than average
\(W < 0\) -> Fewer survivors than expected; center performing worse than average
Provides a volume-adjusted, risk-adjusted measure similar in purpose to RMM.
Example: W-Score Calculation
Let:
\(n = 900\) total patients
\(n_{\text{deaths}} = 40\) deaths
\(\sum P(Survival) = 750.3638\) (sum of predicted survivals)
Step 1: Compute observed survivors
\[
A = n - n_{\text{deaths}}
\]
\[
A = 900 - 40 = 860
\]
Step 2: Define expected survivors
\[
B = \sum P(Survival) = 750.3638
\]
Step 3: Apply W-score formula
\[
W = \frac{A - B}{C} \times 100
\]
Substitute known values:
\[
W = \frac{860 - 750.3638}{900} \times 100
\]
Step 4: Compute W-score
\[
W = \frac{109.6362}{900} \times 100 = 12.18
\]
Step 5: Inference
\(W = 12.18\)
-> The center achieved about 12 more survivors per 100 patients than expected.
Indicates better-than-expected performance after adjusting for patient risk.
W Score is limited
The W Score method is derived from the MTOS study, which was undergirded by linear methods
Divides patients into bins of equal width based on predicted survival probability, \(P(Survival)\).
Assumes that \(P(Survival)\) is evenly distributed.
W Score is limited
Problem:\(P(Survival)\) from logistic regression is not normally distributed — many patients cluster near very high or very low survival probabilities.
Linear bins overrepresent some risk groups and underrepresent others, which can distort observed vs expected comparisons.
Distribution of Predicted Survival
Empirical data show that trauma patients are not evenly distributed across predicted survival probabilities.
Most patients presenting to trauma centers have a very high likelihood of survival.
MTOS Distribution
Ps Range
Proportion of Patients
0.96 – 1.00
0.842
0.91 – 0.95
0.053
0.76 – 0.90
0.052
0.51 – 0.75
0.000
0.26 – 0.50
0.043
0.00 – 0.25
0.010
W-Score Can Be Misleading
The W-score is heavily influenced by the majority of patients with very high \(P(Survival)\) values (for example, \(P(Survival) > 0.8\).
Because most trauma patients are expected to survive, the W-score often reflects performance among the least acute patients, not those at highest risk.
W-Score Can Be Misleading
This means two centers could have identical W-scores even if one performs much better with severely injured patients.
Assumption vs. Reality
The W-score assumes that \(P(Survival)\) values are linearly distributed among patients across the 0–1 range.
However, observed data show that \(P(Survival)\) is highly skewed, with most patients near 1.0.
Therefore, linear bins or evenly spaced \(P(Survival)\) categories overweight low-acuity patients and underweight critical cases.
Take-Home Message on the W Score
W-score alone provides a partial picture of trauma center performance.
For a fair comparison, models such as the Relative Mortality Metric (RMM) use non-linear binning that reflects the true, non-normal \(P(Survival)\) distribution observed in real trauma data.
Relative Mortality Metric (RMM)
Napoli et al. (2017)
The RMM is a risk-adjusted metric that compares observed mortality to predicted mortality.
It accounts for patient-level severity, physiology, and demographics using previously validated coefficients.
Relative Mortality Metric (RMM)
Positive RMM -> higher-than-expected survival.
Negative RMM -> lower-than-expected survival.
Helps benchmark trauma center performance fairly.
Non-Linear Binning: Why It Matters
Because \(P(Survival)\) is skewed, non-linear bins capture the distribution more accurately.
Examples of non-linear binning:
Quantiles (equal number of patients per bin)
Clinically meaningful thresholds (e.g., very high risk vs moderate vs low)
Non-Linear Binning: Why It Matters
This allows fairer comparison of observed vs expected outcomes across risk groups.
Ensures that the benchmarking metrics (RMM, W-score) reflect actual patient risk rather than arbitrary binning.